Retry Strategy

Retry logic is implemented in retry/controller.py and retry/analysis.py.

Retry Decision

Retry only when both are true:

score < threshold
attempt_number < max_attempts

Failure-to-Action Mapping

low Test passing: upgrade model, optionally re-search using failed test messages
low Execution: add MCP server hint + re-search hints
low Syntax validity: upgrade model
low Completeness: upgrade model + re-search

Model Upgrade Paths

Anthropic: claude-sonnet-4-6 -> claude-opus-4-8
OpenAI: gpt-4o-mini -> gpt-4o -> o3
Gemini: gemini-2.5-flash -> gemini-2.5-pro

If already at top model for a provider, model remains unchanged.

Best Attempt Selection

Final response returns the highest score attempt via max(score).